Traditional automated essay scoring methods heavily rely on feature engineering to evaluate and assign scores to essays,which means the quality of the selected features have huge impacts on the performance of such methods.In addition,it also costs a number of labors to manually design the most informative features and requires experts knowledge at most of the time.In this paper,we innovatively propose an adversarial process for learning the relation between an essay and its assigned score,in which we simultaneously train a generative model G and a discriminative model D attempts to distinguish the generated score from the ground truth score without any feature engineering.The experimental results show our approach outperforms several state-of-the-art methods.